A Fault Detection Approach Based on One-Sided Domain Adaptation and Generative Adversarial Networks for Railway Door...

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Domain adaptation is promising for fault detection but faces challenges like ignoring domain-specific decision boundaries and prioritizing source data.
• Real-world implementation of current domain adaptation methods for fault detection is limited due to these practical issues.

Purpose of the Study:

• To propose a novel one-sided domain adaptation approach for accurate and robust fault detection in real-world railway door systems.
• To address the limitations of existing domain adaptation methods, particularly their performance near class boundaries and data utilization.

Main Methods:

• Developed a one-sided domain adaptation technique for fault detection in railway doors.
• Utilized an anomaly detector on label-rich source domain data to generate distinctive source latent features.
• Aligned target domain features towards source latent features in a one-sided manner.

Main Results:

• The proposed method achieved a high F1 score of 97.9%, outperforming alternative methods in accuracy.
• Demonstrated superior robustness against variations in input features compared to existing approaches.
• Successfully bridged the gap between theoretical domain adaptation research and practical industrial applications.

Conclusions:

• The novel one-sided domain adaptation method offers a significant advancement for fault detection in railway systems.
• The approach is applicable to various electro-mechanical actuators and conventional railway components, requiring only motor current signals.
• This research paves the way for more practical and widespread use of domain adaptation in industrial fault detection.